Are Coding Assessments Still Useful When Candidates Have AI Assistants?

The question every hiring leader is asking

If a candidate can paste any standard coding-test question into an LLM and get a near-perfect answer, what is the assessment actually measuring? For old-style algorithmic LeetCode tests, the honest answer is: not much.

But that doesn't mean coding assessments are dead. It means a specific style of coding assessment is dead. The styles that survive — and become more valuable, not less — are different.

What still works

1. Live debugging on unfamiliar code

The candidate is given a small, broken codebase and asked to find and fix the bug. LLMs help less than people assume because the bug is in the interaction between specific files and the candidate has to read the code, not generate it. Tools accelerate good engineers and don't compensate for weak ones.

2. Take-home with walk-through

A 90-minute take-home produces an artifact. A 30-minute walk-through verifies the candidate can reason about it. Together, they remain high-signal even when AI helped with the artifact — because the walk-through tests judgment about the work, which the AI cannot transfer to the candidate.

This is the dominant pattern that emerges from teams who've adapted well: don't fight AI on the artifact, test for it on the explanation.

3. System design

LLMs answer system design questions in the form of an answer but consistently miss the trade-off articulation, the failure-mode reasoning, and the cost awareness that experienced engineers bring. A rubric-anchored system design round with active interviewer pushback remains high-signal.

4. Pair programming on a real task

Collaborative work in real time. The candidate's communication, integration of feedback, and judgment are what's evaluated. AI assistance in the moment is fine — the signal is what they do with it.

What stopped working

1. Algorithmic LeetCode questions

If the question can be solved by pasting it into ChatGPT, you are filtering for who has access to ChatGPT. Retire.

2. Take-homes without walk-through

Pure artifact assessment is unrecoverable. Either add a walk-through or stop using take-homes for high-stakes decisions.

3. MCQ trivia

"What's the time complexity of X" answered in isolation. Easy to look up, easy to AI, doesn't measure judgment. Use only as a screen-stage filter for clear fundamentals gaps, not as decisional signal.

What to add

Process-trace integrity signals

For take-homes, capture keystroke and edit-iteration patterns. ClarityHire does this by default. Doesn't classify good or bad — surfaces patterns inconsistent with hand-written code so the reviewer can probe in the walk-through.

Verbal defense

Make defensibility part of every assessment. The candidate who can use AI tools effectively and explain their own work is the candidate you want. The candidate who pasted without understanding fails the verbal defense regardless of what the artifact looks like.

Realistic problems

Move away from puzzles and toward problems that resemble work. Real problems have ambiguity, context, trade-offs. AI assistants are most helpful on well-specified problems and least helpful on ambiguous ones — exactly the asymmetry you want.

The bigger framing

Coding assessments were never meant to measure "can you write code without help." They were meant to predict job performance. In 2026, job performance includes using AI assistants well. An assessment that pretends those assistants don't exist measures the wrong thing.

The right assessment in 2026 measures: can you produce work, can you explain your work, can you recognise when the AI is wrong, can you handle ambiguity. The first is partly automatable. The other three are not.

Keep the assessments. Redesign them. The signal is still there — it's in different places.